At the bank's Data Fusion conference, Sergey Bezbogov, deputy head of VTB's technology unit, outlined a universal challenge: most AI projects, despite intense interest, fail to move beyond the pilot stage. He identified five systemic barriers that resonate with engineering teams globally.

The first is a strict economic filter. VTB applies a 'lean AI' principle, demanding a clear return on investment (ROI) and a defined payback period, identical to any other tech project. This immediately disqualifies costly pilots with ambiguous value propositions, a significant hurdle for compute-intensive generative models.

This leads directly to the second barrier: infrastructure cost and scarcity. Scaling large language models (LLMs) demands specialized hardware clusters. Often, the economics of running even a highly efficient model are undermined by the price of the required 'iron,' compounded by the global shortage of GPUs and processors.

Third is the reliability problem of model hallucinations. Generative AI producing fabricated facts presents tangible financial and reputational risk. Mitigation requires complex, domain-specific detection systems and cascading architectures—there is no universal fix, adding layers of custom engineering overhead.

Fourth, the data foundation is weak. There's a market-wide deficit of high-quality, curated training data. Creating such datasets is expensive, requiring extensive filtering and expert validation. Within sectors like banking, datasets are often homogenized, stifling innovation. Potential cross-industry data exchanges are blocked by stringent legal restrictions on sensitive information.

Finally, the human factor. Production deployment requires new roles—data annotators and prompt engineers—integrated into cross-functional teams. It also demands a shift in user behavior: operators must learn to frame queries precisely and maintain critical oversight, never fully ceding decision-making authority to the system.

Source: RIA Novosti